Distributed control system remote alarm processing method and device, equipment and medium

Abstract

The application relates to the technical field of industrial control and information, in particular to a distributed control system remote alarm processing method and device, equipment and a medium, the method comprising the following steps: obtaining operation data of a distributed control system; inputting the operation data into a preset mechanism-data hybrid digital twin model to obtain a dynamic health index and corresponding predicted residual life; in the case that the dynamic health index is lower than a preset threshold, performing dynamic correlation retrieval based on a preset four-dimensional knowledge graph to obtain an evolution path; inputting the dynamic health index and the corresponding predicted residual life, and historical cases and the corresponding evolution path into a preset large language model to generate a multi-dimensional natural language description text; deducing multiple candidate schemes in a disposal suggestion to obtain a deduction result; and fusing the natural language description text and the deduction result to generate a decision package. The remote monitoring efficiency and accuracy can be improved.

Description

Technical Field

[0001] This application relates to the technical field of industrial control and information technology, and in particular to remote alarm processing methods, devices, equipment and media for distributed control systems. Background Technology

[0002] Currently, the alarm mechanism of distributed control systems (DCS) is usually based on threshold triggering mode, that is, an alarm is issued when the process parameter exceeds the preset absolute value or rate of change value. With the development of the industrial Internet, when an alarm is issued in this way, the fault has already occurred and the process parameter has deviated from the normal range. For slowly evolving faults, such as equipment wear and tear or decreased thermal efficiency, product quality fluctuations or safety hazards have often already occurred. Moreover, when an alarm is issued, it is usually only an isolated point name, value and description. In order to analyze the cause, the operator needs to manually check multiple heterogeneous systems such as DCS historical database, equipment maintenance records, and environmental monitoring systems, resulting in low information integration efficiency.

[0003] In the aforementioned technologies, once the DCS system malfunctions, there is not only a certain lag, but also low subsequent processing efficiency, which may affect product quality, and in more serious cases, may even pose safety hazards. Summary of the Invention

[0004] This application provides a DCS remote alarm processing method, device, equipment, and medium, which can provide early warning, intelligent diagnosis, and auxiliary decision-making for industrial control processes, thereby improving the efficiency and accuracy of remote monitoring.

[0005] On one hand, embodiments of this application provide a remote alarm processing method for a distributed control system, the method comprising:

[0006] Obtain operational data from the distributed control system;

[0007] The operational data is input into a preset mechanism-data hybrid digital twin model to obtain the dynamic health index and corresponding predicted remaining life of each device in the distributed control system.

[0008] When the dynamic health index is lower than a preset threshold, dynamic association retrieval is performed based on a preset four-dimensional knowledge graph to obtain historical cases and corresponding evolution paths. The preset four-dimensional knowledge graph includes static hierarchical relationships, dynamic real-time operation data, historical maintenance cases, and external environment data.

[0009] The dynamic health index and its corresponding predicted remaining lifespan, along with historical cases and their corresponding evolution paths, are input into a preset large language model. The preset large language model then generates multi-dimensional natural language description text, which includes fault diagnosis, cause analysis, and treatment suggestions.

[0010] Based on the preset mechanism-data hybrid digital twin model, multiple candidate solutions in the proposed treatment are simulated to obtain simulation results. The simulation results include the expected impact curve of each candidate solution on the equipment health index and the extension time of the remaining lifespan.

[0011] The natural language description text and the inference results are fused to generate a decision package, which is then pushed and displayed to the target terminal.

[0012] Optionally, the preset mechanism-data hybrid digital twin model includes a mechanism model and a preset anomaly detection network. The step of inputting the operational data into the preset mechanism-data hybrid digital twin model to obtain the dynamic health index and corresponding predicted remaining life of each device in the distributed control system includes:

[0013] Based on the operational data, a preset mechanism model is driven to run, generating a series of baseline parameters for the device under ideal operating conditions;

[0014] The residual between the running data and the baseline parameter sequence is input into the preset anomaly detection network to extract the deep temporal features of the residual sequence;

[0015] The deep temporal features are nonlinearly fused with the baseline parameter sequence to generate a dynamic health index.

[0016] Based on the time series of the dynamic health index, the probability density distribution of the remaining effective lifespan is output using a time series prediction model to obtain the predicted remaining lifespan.

[0017] Optionally, the step of inputting the residual between the running data and the baseline parameter sequence into the preset anomaly detection network to extract the deep temporal features of the residual sequence includes:

[0018] Construct a variational autoencoder network that includes an encoder and a decoder;

[0019] A long short-term memory network layer is introduced into the encoder; the residual sequence is input into the encoder, and the time dependency of the residual sequence is captured by the long short-term memory network layer and mapped to the normal distribution parameters of the latent space;

[0020] The residual sequence is sampled from the latent space and input into the decoder to reconstruct the residual sequence.

[0021] Calculate the reconstruction error between the reconstructed residual and the original residual;

[0022] The anomaly detection threshold is dynamically adjusted based on the statistical distribution of the reconstruction error.

[0023] When the reconstruction error continues to exceed the dynamic threshold, the encoded vector of the latent space is output as the extracted deep temporal feature.

[0024] Optionally, the step of performing dynamic association retrieval based on a preset four-dimensional knowledge graph to obtain historical cases and corresponding evolution paths includes:

[0025] The current dynamic health index is lower than the preset threshold, and the health index decline trajectory is segmented into linear fits. The slope change points, curvature features and duration of the decline trajectory are extracted as multi-dimensional feature vectors.

[0026] In the historical case library of the preset four-dimensional knowledge graph, the same feature extraction operation is performed on the health index trajectory of each historical case, and the similarity distance between the current trajectory and each historical trajectory is calculated based on the dynamic time warping algorithm;

[0027] The top K historical cases with the smallest similarity distance are selected, and the actual evolution path, fault diagnosis results and handling measures of the historical cases in the subsequent stage of the decline trajectory are extracted as the results of dynamic association retrieval.

[0028] Optionally, the calculation of the similarity distance between the current trajectory and each historical trajectory based on the dynamic time warping algorithm includes:

[0029] The current trajectory and each historical trajectory are length normalized so that the two trajectories have the same time dimension;

[0030] Construct a cost matrix between the two trajectories, wherein the matrix elements of the cost matrix are the absolute values ​​of the difference between the health index values ​​of the two trajectories at corresponding time points;

[0031] A dynamic programming algorithm is used to search for an optimal curved path from the starting point to the ending point in the cost matrix, such that the cumulative cost on the path is minimized, and the minimum cumulative cost is used as the similarity distance between the two trajectories.

[0032] When the time scales of the two trajectories are inconsistent, the slope of the optimal curved path is used to automatically compensate for the time scale difference, so as to achieve accurate matching of equipment trajectories with different decay rates.

[0033] Optionally, the step of inputting the dynamic health index and its corresponding predicted remaining lifespan, along with historical cases and their corresponding evolution paths, into a preset large language model, and generating multi-dimensional natural language descriptive text through the preset large language model, includes:

[0034] The dynamic health index, predicted remaining lifespan, historical cases and evolution paths are encapsulated according to a preset cognitive package data structure to generate structured prompt words. The preset cognitive package data structure includes warning semantics, diagnostic semantics, prediction semantics and suggestion semantics.

[0035] The structured prompt words are input into a preset large language model, and multi-granularity output instructions are set so that the large language model generates first-granularity text suitable for smartwatches, second-granularity text suitable for mobile phones, and third-granularity text suitable for tablets. The first-granularity text, second-granularity text, and third-granularity text maintain semantic consistency but the information density increases progressively.

[0036] Optionally, the step of performing deductions on multiple candidate solutions in the proposed treatment based on the preset mechanism-data hybrid digital twin model to obtain deduction results includes:

[0037] Based on the operation type of the multiple candidate schemes, each candidate scheme is automatically parsed into a corresponding multi-dimensional control instruction set, which includes continuous parameter control instructions, discrete equipment state switching instructions, and time sequence operation instructions.

[0038] The current dynamic health index and corresponding real-time operating status of the device are used as initial conditions and loaded into the preset mechanism-data hybrid digital twin model;

[0039] The digital twin model is driven to run at a first acceleration rate to simulate the dynamic response process of equipment health index, key process parameters and energy consumption indicators after the execution of the multi-dimensional control instruction set;

[0040] During the simulation, when a preset abnormal fluctuation characteristic is detected in the dynamic response process, the digital twin model is automatically triggered to pause the simulation, and the execution timing or parameter amplitude of the multi-dimensional control instruction set is dynamically adjusted based on the current intermediate state to generate an optimized control instruction set.

[0041] Continue driving the digital twin model to run at the second acceleration rate until the complete simulation cycle is completed, generating simulation results that include the health index recovery curve, the trajectory of key parameter changes, and the total amount of resource consumption.

[0042] On the other hand, embodiments of this application provide a remote alarm processing device for a distributed control system, the device comprising:

[0043] The acquisition module is used to acquire the operating data of the distributed control system.

[0044] The input module is used to input the operating data into a preset mechanism-data hybrid digital twin model to obtain the dynamic health index and corresponding predicted remaining life of each device in the distributed control system.

[0045] The retrieval module is used to perform dynamic association retrieval based on a preset four-dimensional knowledge graph when the dynamic health index is lower than a preset threshold, to obtain historical cases and corresponding evolution paths. The preset four-dimensional knowledge graph includes static hierarchical relationships, dynamic real-time operation data, historical maintenance cases, and external environment data.

[0046] The generation module is used to input the dynamic health index and the corresponding predicted remaining lifespan, as well as historical cases and corresponding evolution paths, into a preset large language model, and generate multi-dimensional natural language description text through the preset large language model. The natural language description text includes fault diagnosis, cause analysis and treatment suggestions.

[0047] The simulation module is used to simulate multiple candidate solutions in the proposed treatment based on the preset mechanism-data hybrid digital twin model, and obtain simulation results. The simulation results include the expected impact curve of each candidate solution on the equipment health index and the extension time of the remaining lifespan.

[0048] The fusion module is used to fuse the natural language description text and the inference results to generate a decision package for push and display to the target terminal.

[0049] In another aspect, embodiments of this application provide an electronic device, the device including: a processor and a memory storing computer program instructions;

[0050] When the processor executes the computer program instructions, it implements the remote alarm processing method for the distributed control system as described in the first aspect.

[0051] In another aspect, embodiments of this application provide a computer storage medium storing computer program instructions, which, when executed by a processor, implement the remote alarm processing method for a distributed control system as described in the first aspect.

[0052] The distributed control system remote alarm processing method, apparatus, equipment, and medium of this application embodiment can obtain the dynamic health index and predicted remaining life of distributed control system equipment by introducing a mechanism-data hybrid digital twin model. This enables early warning of potential equipment failures and effectively avoids the lag of traditional threshold alarms. Simultaneously, by combining a four-dimensional knowledge graph and a large language model, isolated operational data is transformed into multi-dimensional natural language description text, providing fault diagnosis, cause analysis, and handling suggestions, significantly improving information integration and understanding efficiency. Furthermore, by using the digital twin model to deduce handling solutions, quantitative evidence is provided for decision-making. The final generated decision package can assist operators in quickly and accurately responding to equipment anomalies, ensuring stable operation of the production process and product quality. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating a remote alarm processing method for a distributed control system provided in an embodiment of this application;

[0054] Figure 2 This is a structural block diagram of a remote alarm processing device for a distributed control system provided in an embodiment of this application;

[0055] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0056] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0057] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0058] To address the problems of existing technologies, this application provides a remote alarm processing method, apparatus, device, and medium for distributed control systems. In this application, by introducing a mechanism-data hybrid digital twin model, the dynamic health index and predicted remaining lifespan of the distributed control system equipment are obtained, enabling early warning of potential equipment failures and effectively avoiding the lag of traditional threshold alarms. Simultaneously, by combining a four-dimensional knowledge graph and a large language model, isolated operational data is transformed into multi-dimensional natural language descriptive text, providing fault diagnosis, cause analysis, and handling suggestions, significantly improving information integration and understanding efficiency. Furthermore, by using the digital twin model to deduce handling solutions, quantitative evidence is provided for decision-making. The final generated decision package can assist operators in quickly and accurately responding to equipment anomalies, ensuring stable operation of the production process and product quality.

[0059] The remote alarm processing method for the distributed control system provided in the embodiments of this application will be introduced first below.

[0060] Figure 1 A flowchart illustrating a remote alarm processing method for a distributed control system according to an embodiment of this application is shown. Figure 1 As shown, the remote alarm processing method for a distributed control system may include S101-S106:

[0061] S101, acquire the operating data of the distributed control system.

[0062] In this embodiment, a distributed control system (DCS) is a widely used control system in industrial production processes. Its main function is to centrally monitor, operate, manage, and distribute the production process. It can include multiple devices, and industrial production is achieved by controlling these devices. Operating data can be collected in real time by various sensors in the system, such as temperature sensors, pressure sensors, and flow sensors, or it can be periodically read from existing control systems or historical databases via data interfaces. For example, a data acquisition module can be configured to retrieve key process parameters, equipment status signals, and other data from the DCS's real-time database at preset time intervals. In addition, some non-real-time data, such as equipment maintenance logs and calibration records, can also be obtained through manual input.

[0063] S102, input the operating data into the preset mechanism-data hybrid digital twin model to obtain the dynamic health index and corresponding predicted remaining life of each device in the distributed control system.

[0064] In this embodiment, the digital twin model can assess the current operating status of a device based on input data using an internally established mathematical or statistical model, and output a quantified health index. For example, the model can analyze the device's vibration data, current data, etc., and characterize the device's health status by calculating the mean, variance, or specific frequency components of these parameters. Simultaneously, the model can also estimate the potential future operating time of the device based on the historical trend of the health index, combined with a simple extrapolation algorithm.

[0065] S103, when the dynamic health index is lower than the preset threshold, perform dynamic association retrieval based on the preset four-dimensional knowledge graph to obtain historical cases and corresponding evolution paths.

[0066] In this embodiment, the pre-defined four-dimensional knowledge graph includes static hierarchical relationships, dynamic real-time operational data, historical maintenance cases, and external environmental data. For example, when the health index of a device drops below the warning line, the system can perform keyword matching or rule-based queries in the knowledge graph based on the device's type, current operating parameters, and other information to find historical fault cases similar to the current situation. The search results can include similar faults that have occurred in the past, their development process, and the final maintenance records.

[0067] S104. Input the dynamic health index and the corresponding predicted remaining lifespan, as well as historical cases and corresponding evolution paths, into the preset large language model, and generate multi-dimensional natural language description text through the preset large language model.

[0068] In this embodiment, the natural language description text includes fault diagnosis, cause analysis, and handling suggestions. For example, a large language model can receive this structured and unstructured information and, based on its training knowledge, generate a report containing the possible fault types of the equipment, the potential causes of the fault, and targeted maintenance or operation suggestions. The model can generate text of varying lengths, such as a short summary or a detailed analysis report, according to preset output length or level of detail requirements.

[0069] It is worth noting that the pre-defined large language model is a natural language processing model based on deep learning. It has been trained on a large amount of text data and is able to understand, generate and process human language, as well as perform complex semantic analysis and reasoning.

[0070] S105, based on the preset mechanism-data hybrid digital twin model, the multiple candidate solutions in the disposal suggestions are deduced to obtain the deduction results.

[0071] In this embodiment, the simulation results include the expected impact curve of each candidate solution on the equipment health index and the extension of remaining lifespan. For multiple disposal suggestions generated by the large language model, the digital twin model can simulate the execution of these suggestions separately. For example, if a suggestion is "reduce motor speed," the digital twin model can simulate how the motor's health index changes after reducing the speed, and how much its remaining lifespan might be extended. This simulation process can use preset simulation parameters and models to quickly evaluate the potential effects of different solutions.

[0072] S106, the natural language description text and the inference results are fused to generate a decision package, which is then pushed and displayed to the target terminal.

[0073] In this embodiment, for example, the diagnostic report generated by the large language model and the solution effect charts derived by the digital twin model can be integrated into a unified interface or document. This decision package can be sent to the operator's workstation, mobile device, or remote monitoring center so that relevant personnel can have a comprehensive understanding of the equipment status, potential risks, and feasible solutions, thereby assisting them in making decisions.

[0074] This embodiment introduces a mechanism-data hybrid digital twin model to obtain the dynamic health index and predicted remaining life of distributed control system equipment, enabling early warning of potential equipment failures and effectively avoiding the lag of traditional threshold alarms. Simultaneously, by combining a four-dimensional knowledge graph and a large language model, isolated operational data is transformed into multi-dimensional natural language descriptions, providing fault diagnosis, cause analysis, and handling suggestions, significantly improving information integration and understanding efficiency. Furthermore, the digital twin model is used to simulate handling solutions, providing quantitative evidence for decision-making. The resulting decision package assists operators in quickly and accurately responding to equipment anomalies, ensuring stable production processes and product quality.

[0075] In other embodiments, the preset mechanism-data hybrid digital twin model includes a mechanism model and a preset anomaly detection network, and S102 may include:

[0076] The system operates based on a pre-defined mechanism model driven by operational data, generating a series of baseline parameters for the equipment under ideal operating conditions.

[0077] The residuals between the running data and the baseline parameter sequence are input into a preset anomaly detection network to extract the deep temporal features of the residual sequence;

[0078] A dynamic health index is generated by nonlinearly fusing deep temporal features with a baseline parameter sequence.

[0079] Based on the time series of the dynamic health index, the probability density distribution of the remaining effective lifespan is output using a time series prediction model to obtain the predicted remaining lifespan.

[0080] In this embodiment, after acquiring the operational data of the distributed control system, a preset mechanistic model is first run based on the operational data to generate a baseline parameter sequence for the equipment under ideal operating conditions. The core of the preset mechanistic-data hybrid digital twin model used in this application lies in the integration of two different types of models: the mechanistic model and the preset anomaly detection network. The mechanistic model is a mathematical model built upon first principles such as physical laws, chemical reactions, and thermodynamics of equipment operation. It can accurately describe the behavior patterns and parameter relationships of the equipment under ideal operating conditions, providing a physically meaningful baseline. The preset anomaly detection network is a data-driven model, typically employing deep learning and machine learning techniques. By learning normal patterns from a large amount of historical operational data, it identifies abnormal behaviors that significantly deviate from normal patterns, and is particularly adept at capturing complex, nonlinear data features and potential fault modes. This hybrid architecture aims to combine the advantages of both models to compensate for the shortcomings of a single model, achieving more comprehensive and accurate equipment status perception and anomaly identification. After acquiring the operational data of the distributed control system, this real-time operational data is first used to drive the preset mechanistic model for simulation operation. Upon receiving actual inputs (such as control commands and environmental parameters), the mechanistic model calculates and outputs a sequence of key parameters that the device should possess under the current input conditions in a perfectly healthy and fault-free ideal state, based on its built-in physical laws and mathematical equations. This baseline parameter sequence represents the device's "health fingerprint" or "theoretical normal behavior," providing a clear reference standard for subsequent anomaly detection. For example, for a pump, the mechanistic model can calculate its outlet flow rate and power consumption under ideal conditions based on data such as its rotational speed and inlet pressure.

[0081] Based on this, the residuals between the operating data and the baseline parameter sequence are input into a pre-defined anomaly detection network to extract deep temporal features from the residual sequence. The residual refers to the difference between the actual operating data of the equipment and the ideal baseline parameter sequence generated by the mechanistic model. By calculating the residuals, the physical regularities of normal equipment operation can be effectively stripped away, thus highlighting abnormal fluctuations caused by equipment degradation, malfunctions, or external interference. These residual sequences are then input into the pre-defined anomaly detection network, which utilizes its powerful pattern recognition capabilities to learn and extract deep temporal features from the residual sequences. These deep temporal features are not merely simple numerical deviations, but also contain complex information such as the evolution trend, periodicity, and mutation patterns of abnormal events, enabling a more refined characterization of the degree and nature of equipment deviations from its normal state.

[0082] Subsequently, the deep temporal features are nonlinearly fused with the baseline parameter sequence to generate a dynamic health index. The generation of the dynamic health index is a crucial step, achieved through the nonlinear fusion of deep temporal features extracted from the residuals with the baseline parameter sequence provided by the mechanistic model. Nonlinear fusion means more than just simple superposition or weighting; it involves using complex mathematical mappings or neural network structures to organically combine the deep temporal features representing actual deviations with the baseline parameter sequence representing ideal states. This fusion method comprehensively considers the physical characteristics of the equipment and abnormal patterns in actual operation, making the generated health index both physically interpretable and sensitively reflecting subtle health changes in the equipment under complex operating conditions, thus providing a comprehensive, dynamic, and accurate equipment health assessment indicator.

[0083] Finally, based on the time series of the dynamic health index, a time-series prediction model is used to output the probability density distribution of the remaining effective life, thus obtaining the predicted remaining life. Once the dynamic health index of the equipment is obtained, it can be used as time series data and input into a pre-defined time-series prediction model. Time-series prediction models, such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer, can learn the historical evolution of the health index and predict its future trends. Furthermore, this model outputs not only a single predicted remaining life value, but also a probability density distribution of the remaining effective life. This means that the prediction result is no longer a definite point, but an interval containing uncertainty, such as "the probability of the equipment failing within the next X days is Y%". This probability density distribution can provide decision-makers with richer and more reliable information, enabling them to better assess risks and formulate more flexible and robust maintenance strategies.

[0084] This application combines a mechanistic model with an anomaly detection network. The mechanistic model provides a physical baseline for equipment under ideal operating conditions, effectively filtering out fluctuations in normal operation. This allows the anomaly detection network to focus on analyzing the residuals between actual operating data and the baseline, thus more sensitively and accurately capturing early, weak anomaly signals and potential degradation trends. This residual analysis, combined with deep temporal feature extraction, can reveal complex failure modes that are difficult to detect using traditional methods. Furthermore, the nonlinear fusion of deep temporal features and baseline parameter sequences ensures that the dynamic health index possesses both physical interpretability and data-driven sensitivity, resulting in more comprehensive and accurate health assessment results. Based on this, the probability density distribution of remaining effective life is output using a time-series prediction model. This not only provides a prediction of equipment lifespan but, more importantly, quantifies the uncertainty of the prediction by providing a probability distribution. This allows decision-makers to more comprehensively assess risks, thereby developing more forward-looking and robust maintenance plans. This avoids misjudgments or delays caused by deviations in a single prediction value, significantly improving the accuracy and reliability of remote alarm processing in distributed control systems.

[0085] In other embodiments, the residual between the running data and the baseline parameter sequence is input into a preset anomaly detection network to extract deep temporal features of the residual sequence, including:

[0086] Construct a variational autoencoder network that includes an encoder and a decoder;

[0087] A long short-term memory network layer is introduced into the encoder; the residual sequence is input into the encoder, and the time dependency of the residual sequence is captured by the long short-term memory network layer and mapped to the normal distribution parameters of the latent space.

[0088] The residual sequence is sampled from the latent space and input into the decoder to reconstruct the residual sequence.

[0089] Calculate the reconstruction error between the reconstructed residual and the original residual;

[0090] The anomaly detection threshold is dynamically adjusted based on the statistical distribution of the reconstruction error.

[0091] When the reconstruction error continues to exceed the dynamic threshold, the encoded vector of the latent space is output as the extracted deep temporal feature.

[0092] In this embodiment, constructing a variational autoencoder network comprising an encoder and a decoder refers to employing a generative model consisting of two main parts: an encoder and a decoder. The encoder is responsible for compressing and mapping the high-dimensional input data (in this case, the residual sequence) to a low-dimensional latent space, while the decoder is responsible for reconstructing the original data from the latent space. By learning the latent representation of the data, the variational autoencoder can capture the inherent structure and distribution characteristics of the data, laying the foundation for subsequent anomaly detection.

[0093] To better handle the temporal characteristics of the residual sequences, a Long Short-Term Memory (LSTM) network layer is introduced into the encoder, and the residual sequences are input into the encoder. LSTM, as a special type of recurrent neural network, has the ability to process and learn long-term dependencies in time-series data. By integrating an LSTM layer into the encoder, the model can effectively capture the complex patterns and dependencies of the residual sequences over time. After processing the residual sequences, the encoder maps them to the normal distribution parameters of the latent space, namely the mean and variance, which together describe the probability distribution of the residual sequences in the latent space.

[0094] Subsequently, samples are taken from the latent space and fed into the decoder to reconstruct the residual sequence. In variational autoencoders, sampling is required from the latent distribution output by the encoder to achieve the mapping from the latent space to the original data space. This sampling operation introduces randomness, which helps the model learn more robust feature representations. The decoder receives the sampled latent vectors and transforms them back into data with the same dimensions and structure as the original residual sequence, thus completing the reconstruction of the residual sequence.

[0095] After reconstruction, it is necessary to calculate the reconstruction error between the reconstructed residual and the original residual. Reconstruction error is a key indicator of the reconstruction capability of a variational autoencoder model, reflecting the accuracy with which the model learns and reconstructs the input data. For normal data, the model should be able to reconstruct it well, resulting in a small reconstruction error; however, for anomalous data, because it deviates from the normal pattern learned by the model, the reconstruction error will increase significantly. Reconstruction error is typically calculated using metrics such as mean squared error (MSE) or mean absolute error (MAE).

[0096] To improve the adaptability and robustness of anomaly detection, this application further dynamically adjusts the anomaly judgment threshold based on the statistical distribution of reconstruction error. Traditional fixed thresholds are prone to false alarms or missed alarms when faced with changes in equipment operating conditions or sensor drift. By statistically analyzing the reconstruction error of historical normal operating data and updating it in real time in conjunction with the recent statistical characteristics of the reconstruction error (such as the mean, variance, and percentiles within a sliding window) or external environmental parameters, dynamic adjustment of the threshold can be achieved, making the anomaly detection system more adaptable.

[0097] Finally, when the reconstruction error continues to exceed the dynamic threshold, the encoded vector from the latent space is output as the extracted deep temporal feature. When the reconstruction error continues to exceed the dynamic threshold, it indicates that the current device operating state may be abnormal. At this point, the encoded vector extracted from the latent space, as a highly compressed and semantically rich "deep temporal feature," can capture the essential information of the abnormal pattern, providing crucial input for subsequent fault diagnosis, classification, or further analysis.

[0098] Through the above technical solution, this application utilizes a variational autoencoder network containing a long short-term memory (LSTM) network layer to effectively capture the complex nonlinear temporal dependencies in the residual sequence of distributed control system (DCS) operating data. Anomaly detection is performed based on reconstruction errors, and combined with dynamically adjusted anomaly thresholds, significantly improving the accuracy and adaptability of anomaly detection, reducing false alarms and false negatives, thus making the generation of dynamic health indices more accurate and reliable. Furthermore, using the encoded vector of the latent space as the deep temporal feature output provides high-dimensional, semantically rich input for subsequent fault diagnosis and cause analysis, helping to understand the nature and evolution trend of anomalies more deeply. The dynamic threshold adjustment mechanism enables the system to adapt to changes in equipment operating conditions and environmental disturbances, improving the system's generalization ability and practicality, and ensuring the timeliness and accuracy of remote alarm processing in the DCS.

[0099] In some other embodiments, S103 may include:

[0100] The current dynamic health index is lower than the preset threshold, and the health index decline trajectory is segmented into linear fits. The slope change points, curvature features and duration of the decline trajectory are extracted as multi-dimensional feature vectors.

[0101] In the historical case library of the pre-set four-dimensional knowledge graph, the same feature extraction operation is performed on the health index trajectory of each historical case, and the similarity distance between the current trajectory and each historical trajectory is calculated based on the dynamic time warping algorithm.

[0102] The top K historical cases with the smallest similarity distance were selected, and the actual evolution path, fault diagnosis results and disposal measures of the historical cases in the subsequent stages of the decline trajectory were extracted as the results of dynamic association retrieval.

[0103] In this embodiment, the health index decline trajectory corresponding to the current triggered dynamic health index falling below a preset threshold is piecewise linearly fitted. This aims to transform the complex, continuous health index decline curve into a set of structured, quantifiable features. Piecewise linear fitting can capture different stages and trend changes in the decline process, such as the turning point from normal operation to slight degradation, and then to accelerated degradation. Slope change points reflect abrupt changes in the decline rate, curvature features describe the degree of bending in the decline trend, and duration provides timescale information about the decline process. Combining these features into a multi-dimensional feature vector provides a high-dimensional, representative data representation for subsequent similarity calculations. In practice, piecewise linear fitting methods such as the Ramer-Douglas-Peucker algorithm or the Peucker algorithm can be used to process the health index decline trajectory. During the fitting process, by setting an appropriate error threshold, the continuous curve is approximated as a series of line segments. Slope change points can be determined by detecting significant differences in the slopes of adjacent line segments. Curvature features can be obtained by calculating the second derivative of the fitted line segment or by using the local radius of curvature of the fitted curve. The duration is directly recorded from the time when the health index first falls below the threshold to the current moment.

[0104] Subsequently, in the historical case library of the pre-defined four-dimensional knowledge graph, the same feature extraction operation is performed on the health index trajectory of each historical case, and the similarity distance between the current trajectory and each historical trajectory is calculated based on the Dynamic Time Warping (DTW) algorithm. To achieve accurate historical case matching, it is necessary to ensure that the current decline trajectory and the historical case trajectory are comparable in the feature space. Therefore, the same piecewise linear fitting and feature extraction operation is also performed on the health index trajectory of each case in the historical case library to generate the corresponding multi-dimensional feature vector. Then, the Dynamic Time Warping (DTW) algorithm is introduced to calculate the similarity distance between the current trajectory and the historical trajectory. The advantage of the DTW algorithm is that it can handle the similarity calculation between sequences of different lengths and time scales, finding the optimal matching path by "bending" the time axis, thus more accurately reflecting the intrinsic similarity between two decline processes, even if their occurrence speed or duration differs. In implementation, firstly, for each historical health index decline trajectory data stored in the historical case library, piecewise linear fitting is performed in the same way as the current trajectory, and slope change points, curvature features, and duration are extracted to form the multi-dimensional feature vector of the historical case. Then, the feature vector of the current decay trajectory and the feature vector of each historical case are used as input to apply the DTW algorithm. The DTW algorithm constructs a cost matrix, where each element represents the distance between two sequences at corresponding points, and then uses dynamic programming to find a path with the minimum cumulative cost, which is the similarity distance between the two trajectories.

[0105] Finally, the top K historical cases with the smallest similarity distance are selected, and their actual evolution paths, fault diagnosis results, and handling measures in the subsequent stages of the decline trajectory are extracted as the results of dynamic association retrieval. After calculating the similarity distance between the current decline trajectory and all historical case trajectories, the most relevant historical cases need to be selected. By selecting the top K historical cases with the smallest similarity distance, it can be ensured that the selected cases are closest to the decline pattern faced by the current equipment. From these highly matched historical cases, their actual evolution paths, final fault diagnosis results, and handling measures taken at that time are extracted. This information is crucial for understanding the possible development trend of the current equipment, predicting potential fault types, and providing effective handling suggestions. In implementation, all historical cases are sorted in ascending order according to their similarity distance with the current trajectory. The top K cases in the sorting results are selected. For these K cases, the complete time series data, recorded fault diagnosis reports, and records of maintenance or intervention operations performed at that time are retrieved and extracted from the preset four-dimensional knowledge graph after the health index decline occurs until the fault occurs or maintenance is completed. The extracted information will serve as the final result of this dynamic association retrieval, providing crucial input for subsequent large language model generation decisions.

[0106] This application's embodiments achieve a refined characterization of the decline pattern by performing piecewise linear fitting on the current and historical health index decline trajectories and extracting multi-dimensional features such as slope change points, curvature characteristics, and duration. Based on this, a dynamic time warping algorithm is used to calculate the similarity distance between trajectories, overcoming the matching difficulties caused by different decline rates and durations. This ensures that even if the decline process differs on a time scale, inherently similar degradation patterns can be accurately identified. Finally, highly relevant historical cases and their subsequent evolution paths, fault diagnosis results, and handling measures are selected, providing high-quality, highly matching contextual information for the large language model. This significantly improves the accuracy and reliability of fault diagnosis, cause analysis, and handling suggestions, thereby enabling more precise guidance for remote alarm handling in distributed control systems and effectively ensuring the stability and safety of equipment operation.

[0107] In some embodiments, the similarity distance between the current trajectory and each historical trajectory is calculated based on a dynamic time warping algorithm, including:

[0108] The current trajectory and each historical trajectory are length normalized so that the two trajectories have the same time dimension;

[0109] Construct a cost matrix between the two trajectories, where each element of the cost matrix is ​​the absolute value of the difference between the health index values ​​of the two trajectories at the corresponding time points.

[0110] The dynamic programming algorithm is used to search for an optimal curved path from the starting point to the ending point in the cost matrix, so as to minimize the cumulative cost on the path, and the minimum cumulative cost is used as the similarity distance between the two trajectories.

[0111] When the time scales of two trajectories are inconsistent, the difference in time scales is automatically compensated by the slope of the optimal curvature path to achieve accurate matching of equipment trajectories with different decay rates.

[0112] Specifically, length normalization is performed on the current trajectory and each historical trajectory to eliminate differences in time length between different health index trajectories. This can be achieved through methods such as resampling, interpolation, or piecewise averaging, downsampling longer trajectories or upsampling shorter ones, ensuring all compared trajectories have a uniform length over time. For example, a standard number of time points can be set, and then linear or spline interpolation can be performed on the original trajectory data to generate a health index sequence at that standard number of time points. This process ensures that subsequent similarity calculations can be performed on a unified time reference, avoiding matching biases caused by inconsistent trajectory lengths.

[0113] When constructing the cost matrix between two trajectories, its elements are the absolute values ​​of the differences between the health index values ​​of the two trajectories at corresponding time points. The cost matrix is ​​a core component of the dynamic time warping algorithm, used to quantify the local differences between the two trajectories at different time points. Specifically, assuming the current trajectory is T1 and the historical trajectory is T2, the element C(i, j) of the cost matrix C represents the absolute value of the difference between the health index value of T1 at time point i and the health index value of T2 at time point j. This absolute value reflects the degree of deviation of the health index states of the two trajectories at a specific time point. By calculating the absolute differences of all possible combinations of time points, a complete cost matrix can be obtained, providing a foundation for subsequently finding the optimal matching path.

[0114] Subsequently, a dynamic programming algorithm is used to search for an optimal curved path from the starting point to the ending point in the cost matrix, minimizing the cumulative cost along the path. This minimum cumulative cost is then used as the similarity distance between the two trajectories. The dynamic programming algorithm is used here to find a path in the cost matrix connecting the starting point (usually the starting point of both trajectories) to the ending point (usually the ending point of both trajectories), where the sum of all cost elements along the path is minimized. This path is called the optimal curved path, which allows the trajectories to non-linearly "bend" or "stretch" along the time axis to find the best alignment. By accumulating the local costs along the path, a global similarity metric, the minimum cumulative cost, can be obtained. The smaller this minimum cumulative cost, the higher the similarity between the two trajectories, even if they have offsets or stretches along the time axis.

[0115] Furthermore, when the time scales of two trajectories are inconsistent, the slope of the optimal curvature path automatically compensates for the time scale difference, achieving accurate matching of equipment trajectories with different decay rates. Time scale inconsistency refers to the possibility that the rate of decay of equipment health indices may differ, causing trajectories to exhibit different rates of change within the same time period. The slope of the optimal curvature path reflects the local alignment of the time axis between the current trajectory and historical trajectories. When the path slope deviates from 1, it indicates that the time axis has been stretched or compressed. For example, a slope greater than 1 indicates that a certain time period of the current trajectory has been stretched to match a shorter time period of the historical trajectory, and vice versa. The optimal curvature path, automatically generated by the dynamic time warping algorithm, inherently includes the ability to compensate for such time scale differences, enabling the finding of the best trajectory morphology match even with different decay rates.

[0116] When calculating the similarity of health index decline trajectories, the current and historical trajectories are first normalized in length to ensure consistency in the time dimension, laying the foundation for subsequent accurate comparison. Next, by constructing a cost matrix and using a dynamic programming algorithm to search for the optimal curved path, the system effectively addresses potential time offsets and local deformations between different trajectories. This allows the similarity calculation to move beyond strict time-point alignment and capture the inherent similarity in trajectory morphology. Crucially, when differences in the decline rate of the equipment health index lead to inconsistent trajectory time scales, the slope of the optimal curved path automatically compensates for these time-scale differences, achieving accurate matching of equipment trajectories with different decline rates. This significantly improves the accuracy and robustness of historical case retrieval, enabling the system to more accurately identify historical fault modes and evolution paths similar to the current equipment state, providing a more reliable basis for subsequent fault diagnosis, root cause analysis, and remedial recommendations.

[0117] In some other embodiments, S104 may include:

[0118] Dynamic health index, predicted remaining lifespan, historical cases and evolution paths are encapsulated according to a preset cognitive package data structure to generate structured prompt words. The preset cognitive package data structure includes warning semantics, diagnostic semantics, prediction semantics and suggestion semantics.

[0119] The structured prompts are input into a preset large language model, and multi-granularity output instructions are set so that the large language model can generate first-granularity text suitable for smartwatches, second-granularity text suitable for mobile phones, and third-granularity text suitable for tablets. The first-granularity text, second-granularity text, and third-granularity text maintain semantic consistency but the information density increases progressively.

[0120] In this embodiment, although a large language model can generate natural language description text containing fault diagnosis, cause analysis, and handling suggestions, the significant differences in screen size, display capabilities, and user interaction habits among different terminals (such as smartwatches, mobile phones, and tablets) are not fully considered when pushing this text to the target terminal for display. This may result in incomplete or overly lengthy information on small-screen terminals, while the amount of information is insufficient on large-screen terminals, thereby affecting the user's efficiency in understanding alarm information and decision-making speed.

[0121] First, dynamic health indices, predicted remaining lifespan, historical cases, and evolution paths are encapsulated according to a pre-defined cognitive package data structure to generate structured prompts. The pre-defined cognitive package data structure is a standardized information organization framework used to uniformly and semantically encapsulate disparate, different types of data (such as numerical health indices, time-series predicted lifespans, and textual descriptions of historical cases). Its core lies in defining four key semantic categories: warning semantics, diagnostic semantics, predictive semantics, and suggestion semantics. Warning semantics describes the anomalies or potential risks in the current equipment state; diagnostic semantics clarifies the possible causes or types of failures; predictive semantics provides trends or remaining lifespan information for the future state of the equipment; and suggestion semantics provides specific handling or maintenance operation guidance. This structured encapsulation ensures that all necessary information is clearly classified and organized before being input into the large language model, laying the foundation for the subsequent generation of high-quality, targeted natural language description text. After encapsulating the information in the pre-defined cognitive package data structure, the system generates one or more structured prompts based on this structured data. These prompts are input instructions specifically designed for large language models. They present the semantic content of the cognitive package, such as warnings, diagnoses, predictions, and suggestions, in a text format that large language models can understand and process efficiently. For example, a prompt might explicitly state, "Please generate a multi-granularity equipment alarm handling report based on the following warning information, diagnostic results, prediction data, and handling suggestions." This structured prompt not only provides all the necessary contextual information but also guides the large language model to focus on specific information points when generating text, ensuring the accuracy and relevance of the output.

[0122] Secondly, structured prompts are input into a preset large language model, and multi-granularity output instructions are set to enable the large language model to generate first-granularity text suitable for smartwatches, second-granularity text suitable for mobile phones, and third-granularity text suitable for tablets. To address the different information display needs of different target terminals, this application introduces multi-granularity output instructions. These instructions are passed to the preset large language model either as part of the structured prompts or as independent parameters. Their function is to explicitly instruct the large language model to adjust the level of detail and information density of the generated text according to the preset granularity level (e.g., for smartwatches, mobile phones, or tablets). By setting such instructions, the large language model can intelligently filter, summarize, or expand information to ensure that the final generated text optimally adapts to the display capabilities of the target terminal and the user's need for information depth. The first-granularity text, second-granularity text, and third-granularity text maintain semantic consistency but with progressively increasing information density. The first-granularity text is designed for terminals with limited display areas, such as smartwatches, and is characterized by highly refined information, typically containing only the most core warning information and emergency handling suggestions to ensure that users can immediately grasp key situations during rapid browsing. The second-level granularity text is suitable for medium-sized terminals such as mobile phones. Building upon the first-level granularity text, it adds a summary of the fault diagnosis, a brief explanation of the cause analysis, and more specific handling suggestions. With moderate information density, it facilitates initial judgment and operation for users in mobile scenarios. The third-level granularity text is designed specifically for large-screen terminals such as tablets. It has the highest information density, including detailed fault diagnosis reports, multi-dimensional cause analysis, in-depth comparisons of historical cases, and potentially detailed handling steps, potential risk assessments, and references to relevant technical documents, aiming to provide users with comprehensive decision support information. Despite the differences in information density between the different granularities, they maintain strict semantic consistency. This means that regardless of the terminal through which the user obtains the information, the core warning content, diagnostic conclusions, prediction results, and handling suggestions are consistent and not contradictory. The progressively increasing information density is achieved by the large language model summarizing, simplifying, or detailing the information to varying degrees according to granularity instructions, without altering the core semantics, when generating text.

[0123] When generating natural language descriptive text, the input information is structurally encapsulated through a pre-defined cognitive package data structure. Combined with multi-granularity output instructions, this guides the large language model to generate text with different information densities based on the characteristics of the target terminal. This ensures that information is presented accurately in the most suitable form for the terminal. For example, smartwatch users can quickly obtain core warnings and emergency suggestions, avoiding information overload; mobile phone users can obtain more detailed diagnostic and treatment plans; and tablet users can obtain comprehensive analysis reports and decision support. This information adaptability significantly improves the user experience and reduces the cognitive burden on users processing information on different terminals. Simultaneously, customized information is provided for different scenarios and roles (such as field operators, maintenance engineers, and managers), thereby accelerating fault response and decision-making processes. Furthermore, strict semantic consistency is maintained between texts of different granularities, avoiding misjudgments or misoperations caused by information differences, further improving the accuracy and reliability of decision-making. Through structured prompts and multi-granularity instructions, this application more effectively utilizes the text generation capabilities of the large language model, making its output more targeted and practical, thereby maximizing its effectiveness in remote alarm processing of distributed control systems.

[0124] In some other embodiments, S105 may include:

[0125] Based on the operation types of multiple candidate schemes, each candidate scheme is automatically parsed into a corresponding multi-dimensional control instruction set. The multi-dimensional control instruction set includes continuous parameter control instructions, discrete equipment state switching instructions, and time sequence operation instructions.

[0126] The current dynamic health index and corresponding real-time operating status of the device are used as initial conditions and loaded into the preset mechanism-data hybrid digital twin model;

[0127] Drive the digital twin model to run at the first acceleration rate to simulate the dynamic response process of equipment health index, key process parameters and energy consumption indicators after executing multi-dimensional control instruction set;

[0128] During the simulation, when a preset abnormal fluctuation characteristic is detected in the dynamic response process, the digital twin model is automatically triggered to pause the simulation and dynamically adjust the execution timing or parameter amplitude of the multi-dimensional control instruction set based on the current intermediate state to generate an optimized control instruction set.

[0129] Continue driving the digital twin model to run at the second acceleration rate until the complete simulation cycle is completed, generating simulation results that include the health index recovery curve, the trajectory of key parameter changes, and the total amount of resource consumption.

[0130] Specifically, firstly, based on the operation types of multiple candidate solutions, the system can automatically parse each candidate solution into a corresponding multi-dimensional control instruction set. This multi-dimensional control instruction set is a structured, executable set of instructions, comprising continuous parameter control instructions, discrete equipment state switching instructions, and time-series operation sequence instructions. Specifically, continuous parameter control instructions are used to adjust continuous variables in equipment operation, such as setting temperature, pressure, flow rate, or speed; these parameters can change smoothly within a preset range. Discrete equipment state switching instructions change the discrete states of the equipment, such as starting or stopping the equipment, switching operating modes, or opening and closing valves. Time-series operation sequence instructions ensure that a series of operations are executed according to a specific logical order and time intervals to handle complex handling processes. This automatic parsing mechanism ensures that the handling recommendations can be accurately understood and executed by the digital twin model.

[0131] Subsequently, the current dynamic health index and corresponding real-time operating status of the equipment are used as initial conditions and loaded into a preset mechanism-data hybrid digital twin model. This aims to ensure that the simulation process starts from a point highly consistent with the current state of the real equipment, thereby improving the accuracy of the simulation results and their practical guiding significance. The dynamic health index provides a quantitative assessment of the overall health status of the equipment, while the real-time operating status includes detailed information such as all key sensor data, controller outputs, and equipment modes.

[0132] Based on this, the digital twin model is driven to run at the first acceleration rate to simulate the dynamic response of equipment health index, key process parameters, and energy consumption indicators after the execution of a multi-dimensional control command set. The first acceleration rate allows the system to perform rapid simulations at a speed far exceeding actual time, in order to preliminarily assess the potential effects and risks of the treatment plan. During the simulation, the digital twin model predicts the changing trends of equipment health index, the fluctuations of key process parameters (such as temperature, pressure, vibration, etc.), and the dynamic response of energy consumption indicators (such as electricity consumption, material consumption, etc.) based on the input control command set.

[0133] Crucially, during the simulation, when pre-defined abnormal fluctuations are detected in the dynamic response process, the system automatically triggers a pause in the digital twin model simulation. These pre-defined abnormal fluctuations can include a sharp drop in health indices, key process parameters exceeding safe limits, and abnormal spikes in energy consumption. Once these anomalies are detected, the simulation process does not continue blindly but is immediately paused. The system dynamically adjusts the execution timing or parameter amplitude of the multi-dimensional control command set based on the current intermediate state, generating an optimized control command set. This means the system can analyze the equipment state at the time of the anomaly and use optimization algorithms or expert rules to adjust the original control command set in real time, such as changing the execution order of commands or modifying parameter settings, to avoid or mitigate the anomaly, thereby generating a safer and more effective optimized control command set.

[0134] Finally, the digital twin model continues to run at the second acceleration rate until the complete simulation cycle is finished, generating simulation results that include the health index recovery curve, the trajectory of key parameter changes, and the total resource consumption. After instruction set optimization, the digital twin model will continue simulation at the second acceleration rate to complete the simulation cycle of the entire disposal plan. The final simulation results will provide a detailed picture of how the equipment health index recovers from the abnormal state, how key process parameters stabilize within the target range, and the total resource consumption required for the entire disposal process, providing comprehensive and reliable data support for the final decision.

[0135] During simulation, this application transforms abstract handling suggestions into an executable multi-dimensional control command set, using the current actual state of the equipment as the starting point, making the simulation process more realistic. More importantly, it introduces a real-time detection and dynamic adjustment mechanism for abnormal fluctuation characteristics during the simulation, avoiding potential risks caused by imperfect initial handling plans and ensuring the robustness and safety of the simulation process. By dynamically optimizing the execution timing or parameter amplitude of the control command set, this application can generate more reasonable and effective handling plans, significantly improving the reliability and practicality of the decision package, thus effectively solving the limitations of traditional simulation methods in lacking real-time adaptability and optimization capabilities when facing complex operating conditions. Furthermore, the phased accelerated simulation mechanism also improves evaluation efficiency, providing more accurate and reliable decision support for remote alarm processing of distributed control systems.

[0136] In this embodiment, to improve the accuracy and optimization effect of the simulation results, when a preset abnormal fluctuation characteristic is detected in the dynamic response process, the digital twin model is automatically triggered to pause the simulation, and the execution timing or parameter amplitude of the multi-dimensional control instruction set is dynamically adjusted based on the current intermediate state, including:

[0137] During the simulation, the rate of change of the equipment health index and the fluctuation range of key process parameters are monitored in real time.

[0138] When the instantaneous decline rate of the health index exceeds the first threshold, or the fluctuation range of the key process parameters exceeds the second threshold, it is determined that an abnormal fluctuation event has occurred in the simulation process.

[0139] Freeze the device state at the current simulation moment as an intermediate state, and call the preset inference path backtracking algorithm to revert to the last stable moment before the abnormal fluctuation event occurred;

[0140] Based on the deviation between the intermediate state and the target desired state, the adjustment amount of the control command is recalculated through an optimization algorithm;

[0141] The adjusted control instruction set is concatenated with the remaining unexecuted original instructions to generate a deduced correction path, and simulation continues based on the correction path.

[0142] Specifically, the system collects data on the equipment health index and key process parameters such as temperature, pressure, flow rate, and vibration output by the digital twin model at a preset sampling frequency. The rate of change of the health index can be obtained by calculating the ratio of the difference in the health index at adjacent time points to the time interval, or by using methods such as sliding window averaging or exponential smoothing. The fluctuation amplitude of key process parameters can be measured by calculating their standard deviation, peak-to-peak value, or root mean square value within a certain time window. This real-time, fine-grained monitoring mechanism aims to provide richer status information than a single instantaneous value, so as to identify the early signs of anomalies more accurately.

[0143] Based on the aforementioned real-time monitoring results, this application establishes clear anomaly judgment criteria. When the instantaneous rate of decline of the equipment health index exceeds a preset first threshold, it indicates that the equipment health condition is rapidly deteriorating; or, when the fluctuation amplitude of any key process parameter exceeds a preset second threshold, it indicates a significant deviation in equipment operational stability. These two thresholds can be set according to the equipment type, historical operating data, expert experience, or safety regulations. Once either condition is met, the system determines that an abnormal fluctuation event has occurred in the current simulation process, requiring intervention. This judgment method based on the rate of change and fluctuation amplitude can effectively avoid misjudgment due to instantaneous noise or slight fluctuations, while also promptly detecting potential serious anomalies.

[0144] Once an abnormal fluctuation event is detected, the system immediately freezes all relevant status parameters of the equipment at the current simulation moment, saving it as an "intermediate state." This intermediate state is a complete system snapshot, containing all necessary information such as equipment health indices, instantaneous values ​​of all key process parameters, internal variables, and the execution progress of control commands. The purpose of freezing the intermediate state is to accurately record the system context at the time of the anomaly, providing an accurate and reliable starting point for subsequent analysis, backtracking, and adjustments.

[0145] To more effectively correct anomalies, this application introduces a backtracking algorithm for predictive paths. Upon detecting an abnormal fluctuation event, this algorithm does not directly adjust based on a frozen intermediate state. Instead, it analyzes simulated historical data to identify the last "stable moment" before the abnormal fluctuation event occurred. The determination of the stable moment can be based on indicators such as the rate of change of the health index and the fluctuation amplitude of key process parameters, ensuring that the system is within a normal or controllable range at that moment. By loading the equipment state at this stable moment, the backtracking algorithm allows subsequent adjustments to begin from a baseline unaffected by the anomaly, thereby avoiding errors caused by the cumulative effect of anomalies and improving the effectiveness and accuracy of adjustments.

[0146] After reverting to a stable point, this application recalculates the adjustment amount of the control command based on the deviation between the equipment state at that stable point (i.e., the "intermediate state" after reverting) and the preset "target desired state." The target desired state could be the equipment health index recovering to a certain expected safety level, key process parameters stabilizing within the optimal operating range, or the remaining lifespan reaching a certain extension target. The optimization algorithm (e.g., using model predictive control, reinforcement learning, or heuristic search methods) calculates the optimal adjustment amount of the control command with the goal of minimizing the deviation between the current state and the target desired state. This adjustment based on the optimization algorithm ensures the scientific validity and effectiveness of the correction scheme.

[0147] After calculating the adjustment amounts of the control commands, this application integrates these adjustments into the original multi-dimensional control command set to form an "adjusted control command set." This adjusted command set replaces or modifies the original commands starting from the backtracking point. Simultaneously, the remaining unexecuted original commands (i.e., commands after the backtracking point but not yet simulated) are logically concatenated with the adjusted command set to generate a complete "deduction correction path." The digital twin model will then continue simulation from the backtracked stable moment, following this new correction path, until the entire deduction cycle is completed. This approach ensures the continuity of the deduction process and the integrity of the correction scheme.

[0148] This application monitors the rate of change of equipment health indices and the fluctuation range of key process parameters in real time, and sets clear thresholds for anomaly detection. This allows the system to promptly detect the initial signs or early development trends of anomalies, rather than intervening only when the anomaly has significantly impacted the system. Furthermore, when an anomaly occurs, this application no longer passively adjusts based solely on the current anomalous state. Instead, it invokes a pre-defined backtracking algorithm to revert the simulation to the last stable moment before the anomalous fluctuation event. This mechanism effectively avoids errors caused by the cumulative effect of anomalies, allowing subsequent adjustments to control commands to begin from an earlier and more stable baseline, thus significantly improving the timeliness and effectiveness of adjustments. Based on this, and considering the deviation between the backtracked intermediate state and the target desired state, an optimized algorithm recalculates the adjustment amount of the control commands, ensuring the scientific validity and accuracy of the correction scheme. Finally, the adjusted control instruction set is concatenated with the remaining unexecuted original instructions to generate a deduced correction path. Simulation continues based on this path, which significantly improves the accuracy, reliability, and robustness of the optimization scheme of the digital twin model deduction results. This makes the generated disposal suggestions more instructive and operable, effectively avoiding the problem of poor overall optimization effect caused by local anomalies in the deduction process. It provides more intelligent and reliable decision support for remote alarm processing of distributed control systems.

[0149] Based on the remote alarm processing method for distributed control systems provided in the above embodiments, this application also provides specific implementation methods for the remote alarm processing device for distributed control systems. Please refer to the following embodiments.

[0150] First see Figure 2 The distributed control system remote alarm processing device 200 provided in this application embodiment may include:

[0151] The acquisition module 201 is used to acquire the operating data of the distributed control system;

[0152] Input module 202 is used to input operating data into a preset mechanism-data hybrid digital twin model to obtain the dynamic health index and corresponding predicted remaining life of each device in the distributed control system;

[0153] The retrieval module 203 is used to perform dynamic association retrieval based on a preset four-dimensional knowledge graph when the dynamic health index is lower than a preset threshold, to obtain historical cases and corresponding evolution paths. The preset four-dimensional knowledge graph includes static hierarchical relationships, dynamic real-time operation data, historical maintenance cases and external environment data.

[0154] The generation module 204 is used to input the dynamic health index and the corresponding predicted remaining lifespan, as well as historical cases and corresponding evolution paths into the preset large language model, and generate multi-dimensional natural language description text through the preset large language model. The natural language description text includes fault diagnosis, cause analysis and treatment suggestions.

[0155] The simulation module 205 is used to simulate multiple candidate solutions in the disposal suggestions based on a preset mechanism-data hybrid digital twin model to obtain simulation results. The simulation results include the expected impact curve of each candidate solution on the equipment health index and the extension time of the remaining lifespan.

[0156] The fusion module 206 is used to fuse the natural language description text and the inference results to generate a decision package for push and display to the target terminal.

[0157] As an alternative implementation, the input module 202 can also be used for:

[0158] The system operates based on a pre-defined mechanism model driven by operational data, generating a series of baseline parameters for the equipment under ideal operating conditions.

[0159] The residuals between the running data and the baseline parameter sequence are input into a preset anomaly detection network to extract the deep temporal features of the residual sequence;

[0160] A dynamic health index is generated by nonlinearly fusing deep temporal features with a baseline parameter sequence.

[0161] Based on the time series of the dynamic health index, the probability density distribution of the remaining effective lifespan is output using a time series prediction model to obtain the predicted remaining lifespan.

[0162] As an alternative implementation, the input module 202 can also be used for:

[0163] Construct a variational autoencoder network that includes an encoder and a decoder;

[0164] A long short-term memory network layer is introduced into the encoder; the residual sequence is input into the encoder, and the time dependency of the residual sequence is captured by the long short-term memory network layer and mapped to the normal distribution parameters of the latent space.

[0165] The residual sequence is sampled from the latent space and input into the decoder to reconstruct the residual sequence.

[0166] Calculate the reconstruction error between the reconstructed residual and the original residual;

[0167] The anomaly detection threshold is dynamically adjusted based on the statistical distribution of the reconstruction error.

[0168] When the reconstruction error continues to exceed the dynamic threshold, the encoded vector of the latent space is output as the extracted deep temporal feature.

[0169] As an alternative implementation, the retrieval module 203 can also be used for:

[0170] The current dynamic health index is lower than the preset threshold, and the health index decline trajectory is segmented into linear fits. The slope change points, curvature features and duration of the decline trajectory are extracted as multi-dimensional feature vectors.

[0171] In the historical case library of the pre-set four-dimensional knowledge graph, the same feature extraction operation is performed on the health index trajectory of each historical case, and the similarity distance between the current trajectory and each historical trajectory is calculated based on the dynamic time warping algorithm.

[0172] The top K historical cases with the smallest similarity distance were selected, and the actual evolution path, fault diagnosis results and disposal measures of the historical cases in the subsequent stages of the decline trajectory were extracted as the results of dynamic association retrieval.

[0173] As an alternative implementation, the retrieval module 203 can also be used for:

[0174] The current trajectory and each historical trajectory are length normalized so that the two trajectories have the same time dimension;

[0175] Construct a cost matrix between the two trajectories, where each element of the cost matrix is ​​the absolute value of the difference between the health index values ​​of the two trajectories at the corresponding time points.

[0176] The dynamic programming algorithm is used to search for an optimal curved path from the starting point to the ending point in the cost matrix, so as to minimize the cumulative cost on the path, and the minimum cumulative cost is used as the similarity distance between the two trajectories.

[0177] When the time scales of two trajectories are inconsistent, the difference in time scales is automatically compensated by the slope of the optimal curvature path to achieve accurate matching of equipment trajectories with different decay rates.

[0178] As an alternative implementation, the generation module 204 can also be used for:

[0179] Dynamic health index, predicted remaining lifespan, historical cases and evolution paths are encapsulated according to a preset cognitive package data structure to generate structured prompt words. The preset cognitive package data structure includes warning semantics, diagnostic semantics, prediction semantics and suggestion semantics.

[0180] The structured prompts are input into a preset large language model, and multi-granularity output instructions are set so that the large language model can generate first-granularity text suitable for smartwatches, second-granularity text suitable for mobile phones, and third-granularity text suitable for tablets. The first-granularity text, second-granularity text, and third-granularity text maintain semantic consistency but the information density increases progressively.

[0181] As an alternative implementation, the deduction module 205 can also be used for:

[0182] Based on the operation types of multiple candidate schemes, each candidate scheme is automatically parsed into a corresponding multi-dimensional control instruction set. The multi-dimensional control instruction set includes continuous parameter control instructions, discrete equipment state switching instructions, and time sequence operation instructions.

[0183] The current dynamic health index and corresponding real-time operating status of the device are used as initial conditions and loaded into the preset mechanism-data hybrid digital twin model;

[0184] Drive the digital twin model to run at the first acceleration rate to simulate the dynamic response process of equipment health index, key process parameters and energy consumption indicators after executing multi-dimensional control instruction set;

[0185] During the simulation, when a preset abnormal fluctuation characteristic is detected in the dynamic response process, the digital twin model is automatically triggered to pause the simulation and dynamically adjust the execution timing or parameter amplitude of the multi-dimensional control instruction set based on the current intermediate state to generate an optimized control instruction set.

[0186] Continue driving the digital twin model to run at the second acceleration rate until the complete simulation cycle is completed, generating simulation results that include the health index recovery curve, the trajectory of key parameter changes, and the total amount of resource consumption.

[0187] Figure 3 A schematic diagram of the hardware structure of the electronic device provided in an embodiment of this application is shown.

[0188] An electronic device may include a processor 301 and a memory 302 storing computer program instructions.

[0189] Specifically, the processor 301 may include a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0190] Memory 302 may include mass storage for data or instructions. For example, and not limitingly, memory 302 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. In one instance, memory 302 may include removable or non-removable (or fixed) media, or memory 302 may be non-volatile solid-state memory. Memory 302 may be internal or external to the integrated gateway disaster recovery device.

[0191] In one instance, memory 302 may be read-only memory (ROM). In one instance, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically rewritable ROM (EAROM), or flash memory, or a combination of two or more of these.

[0192] Memory 302 may include read-only memory (ROM), random access memory (RAM), disk storage media device, optical storage media device, flash memory device, electrical, optical, or other physical / tangible memory storage device. Therefore, typically, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the distributed control system remote alarm processing method according to the first aspect of this disclosure.

[0193] The processor 301 reads and executes computer program instructions stored in the memory 302 to achieve... Figure 1 A remote alarm processing method for a distributed control system is shown in the embodiment.

[0194] In one example, the electronic device may also include a communication interface 303 and a bus 304. For example, Figure 3 As shown, the processor 301, memory 302, and communication interface 303 are connected through bus 304 and complete communication with each other.

[0195] The communication interface 303 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0196] Bus 304 includes hardware, software, or both, that couples components of an electronic device together. For example, and not as a limitation, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or a combination of two or more of these. Where appropriate, bus 304 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, this application contemplates any suitable bus or interconnect.

[0197] This electronic device can execute the remote alarm processing method of the distributed control system in the embodiments of this application, thereby achieving a combination of Figures 1-3 The method and apparatus for remote alarm processing in a distributed control system are described.

[0198] Furthermore, in conjunction with the remote alarm processing method for distributed control systems in the above embodiments, this application embodiment can provide a computer storage medium for implementation. The computer storage medium stores computer program instructions; when these computer program instructions are executed by a processor, they implement any of the remote alarm processing methods for distributed control systems in the above embodiments.

[0199] In an optional embodiment, in conjunction with the remote alarm processing method for distributed control systems in the above embodiments, this application embodiment can provide a computer program product to implement it. The instructions in the computer program product are executed by the processor of the electronic device, enabling the electronic device to implement any of the remote alarm processing methods for distributed control systems in the above embodiments.

[0200] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0201] The functional blocks shown in the above block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0202] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0203] The aspects of this disclosure have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by special-purpose hardware performing the specified functions or actions, or can be implemented by a combination of special-purpose hardware and computer instructions.

[0204] The above description is merely a specific implementation of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A remote alarm processing method for a distributed control system, characterized in that, include: Obtain operational data from the distributed control system; The operational data is input into a preset mechanism-data hybrid digital twin model to obtain the dynamic health index and corresponding predicted remaining life of each device in the distributed control system. When the dynamic health index is lower than a preset threshold, dynamic association retrieval is performed based on a preset four-dimensional knowledge graph to obtain historical cases and corresponding evolution paths. The preset four-dimensional knowledge graph includes static hierarchical relationships, dynamic real-time operation data, historical maintenance cases, and external environment data. The dynamic health index and its corresponding predicted remaining lifespan, along with historical cases and their corresponding evolution paths, are input into a preset large language model. The preset large language model then generates multi-dimensional natural language description text, which includes fault diagnosis, cause analysis, and treatment suggestions. Based on the preset mechanism-data hybrid digital twin model, multiple candidate solutions in the proposed treatment are simulated to obtain simulation results. The simulation results include the expected impact curve of each candidate solution on the equipment health index and the extension time of the remaining lifespan. The natural language description text and the inference results are fused to generate a decision package, which is then pushed and displayed to the target terminal.

2. The method according to claim 1, characterized in that, The preset mechanism-data hybrid digital twin model includes a mechanism model and a preset anomaly detection network. The step of inputting the operational data into the preset mechanism-data hybrid digital twin model to obtain the dynamic health index and corresponding predicted remaining lifespan of each device in the distributed control system includes: Based on the operational data, a preset mechanism model is driven to run, generating a series of baseline parameters for the device under ideal operating conditions; The residual between the running data and the baseline parameter sequence is input into the preset anomaly detection network to extract the deep temporal features of the residual sequence; The deep temporal features are nonlinearly fused with the baseline parameter sequence to generate a dynamic health index. Based on the time series of the dynamic health index, the probability density distribution of the remaining effective lifespan is output using a time series prediction model to obtain the predicted remaining lifespan.

3. The method according to claim 2, characterized in that, The step of inputting the residual between the running data and the baseline parameter sequence into the preset anomaly detection network and extracting the deep temporal features of the residual sequence includes: Construct a variational autoencoder network that includes an encoder and a decoder; A long short-term memory network layer is introduced into the encoder; the residual sequence is input into the encoder, and the time dependency of the residual sequence is captured by the long short-term memory network layer and mapped to the normal distribution parameters of the latent space; The residual sequence is sampled from the latent space and input into the decoder to reconstruct the residual sequence. Calculate the reconstruction error between the reconstructed residual and the original residual; The anomaly detection threshold is dynamically adjusted based on the statistical distribution of the reconstruction error. When the reconstruction error continues to exceed the dynamic threshold, the encoded vector of the latent space is output as the extracted deep temporal feature.

4. The method according to claim 1, characterized in that, The dynamic association retrieval based on a preset four-dimensional knowledge graph to obtain historical cases and corresponding evolution paths includes: The current dynamic health index is lower than the preset threshold, and the health index decline trajectory is segmented into linear fits. The slope change points, curvature features and duration of the decline trajectory are extracted as multi-dimensional feature vectors. In the historical case library of the preset four-dimensional knowledge graph, the same feature extraction operation is performed on the health index trajectory of each historical case, and the similarity distance between the current trajectory and each historical trajectory is calculated based on the dynamic time warping algorithm; The top K historical cases with the smallest similarity distance are selected, and the actual evolution path, fault diagnosis results and handling measures of the historical cases in the subsequent stage of the decline trajectory are extracted as the results of dynamic association retrieval.

5. The method according to claim 4, characterized in that, The calculation of the similarity distance between the current trajectory and each historical trajectory based on the dynamic time warping algorithm includes: The current trajectory and each historical trajectory are length normalized so that the two trajectories have the same time dimension; Construct a cost matrix between the two trajectories, wherein the matrix elements of the cost matrix are the absolute values ​​of the difference between the health index values ​​of the two trajectories at corresponding time points; A dynamic programming algorithm is used to search for an optimal curved path from the starting point to the ending point in the cost matrix, such that the cumulative cost on the path is minimized, and the minimum cumulative cost is used as the similarity distance between the two trajectories. When the time scales of the two trajectories are inconsistent, the slope of the optimal curved path is used to automatically compensate for the time scale difference, so as to achieve accurate matching of equipment trajectories with different decay rates.

6. The method according to claim 1, characterized in that, The process involves inputting the dynamic health index and its corresponding predicted life expectancy, along with historical cases and their corresponding evolution paths, into a pre-defined large language model. The pre-defined large language model then generates multi-dimensional natural language descriptive text, including: The dynamic health index, predicted remaining lifespan, historical cases and evolution paths are encapsulated according to a preset cognitive package data structure to generate structured prompt words. The preset cognitive package data structure includes warning semantics, diagnostic semantics, prediction semantics and suggestion semantics. The structured prompt words are input into a preset large language model, and multi-granularity output instructions are set so that the large language model generates first-granularity text suitable for smartwatches, second-granularity text suitable for mobile phones, and third-granularity text suitable for tablets. The first-granularity text, second-granularity text, and third-granularity text maintain semantic consistency but the information density increases progressively.

7. The method according to claim 1, characterized in that, The deduction of multiple candidate solutions in the proposed treatment based on the preset mechanism-data hybrid digital twin model yields the following results: Based on the operation type of the multiple candidate schemes, each candidate scheme is automatically parsed into a corresponding multi-dimensional control instruction set, which includes continuous parameter control instructions, discrete equipment state switching instructions, and time sequence operation instructions. The current device's dynamic health index and corresponding real-time operating status are used as initial conditions and loaded into the preset mechanism-data hybrid digital twin model; The digital twin model is driven to run at a first acceleration rate to simulate the dynamic response process of equipment health index, key process parameters and energy consumption indicators after the execution of the multi-dimensional control instruction set; During the simulation, when a preset abnormal fluctuation characteristic is detected in the dynamic response process, the digital twin model is automatically triggered to pause the simulation, and the execution timing or parameter amplitude of the multi-dimensional control instruction set is dynamically adjusted based on the current intermediate state to generate an optimized control instruction set. Continue driving the digital twin model to run at the second acceleration rate until the complete simulation cycle is completed, generating simulation results that include the health index recovery curve, the trajectory of key parameter changes, and the total amount of resource consumption.

8. A remote alarm processing device for a distributed control system, characterized in that, The device includes: The acquisition module is used to acquire the operating data of the distributed control system. The input module is used to input the operating data into a preset mechanism-data hybrid digital twin model to obtain the dynamic health index and corresponding predicted remaining life of each device in the distributed control system. The retrieval module is used to perform dynamic association retrieval based on a preset four-dimensional knowledge graph when the dynamic health index is lower than a preset threshold, to obtain historical cases and corresponding evolution paths. The preset four-dimensional knowledge graph includes static hierarchical relationships, dynamic real-time operation data, historical maintenance cases, and external environment data. The generation module is used to input the dynamic health index and the corresponding predicted remaining lifespan, as well as historical cases and corresponding evolution paths, into a preset large language model, and generate multi-dimensional natural language description text through the preset large language model. The natural language description text includes fault diagnosis, cause analysis and treatment suggestions. The simulation module is used to simulate multiple candidate solutions in the proposed treatment based on the preset mechanism-data hybrid digital twin model, and obtain simulation results. The simulation results include the expected impact curve of each candidate solution on the equipment health index and the extension time of the remaining lifespan. The fusion module is used to fuse the natural language description text and the inference results to generate a decision package for push and display to the target terminal.

9. An electronic device, characterized in that, The device includes: a processor and a memory storing computer program instructions; When the processor executes the computer program instructions, it implements the remote alarm processing method for the distributed control system as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the remote alarm processing method for a distributed control system as described in any one of claims 1-7.

Images